From 128cb114bdac6b8cf41a9d689f0a573a2e27eced Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 5 Sep 2026 20:45:47 -0700
Subject: [PATCH] style: trim comments on batch retrieve model group stamp

---
 litellm/router.py                 |  7 ++-----
 tests/test_litellm/test_router.py | 16 ----------------
 2 files changed, 2 insertions(+), 21 deletions(-)

diff --git a/litellm/router.py b/litellm/router.py
index c6c25b6be17..20c9018abee 100644
--- a/litellm/router.py
+++ b/litellm/router.py
@@ -6175,11 +6175,8 @@ class Router:
                         kwargs=new_kwargs,
                         function_name="aretrieve_batch",
                     )
-                    ## STAMP THE MODEL GROUP FOR SPEND TRACKING ##
-                    # A batch is retrieved by id, so the request carries no model group of its
-                    # own - only the deployment that answered knows it. Batch token usage lands
-                    # on this retrieve call (the provider reports counts once the job finishes),
-                    # so without this the tokens are logged under an empty model_group.
+                    # A batch is retrieved by id, so only the deployment that answered knows the
+                    # group, and batch token usage is logged on this retrieve call.
                     model_group: Final = requested_model_group or model_name["model_name"]
                     new_kwargs[metadata_variable_name].setdefault("model_group", model_group)
                     new_kwargs.pop("custom_llm_provider", None)
diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py
index 9b146092927..258d1c973d0 100644
--- a/tests/test_litellm/test_router.py
+++ b/tests/test_litellm/test_router.py
@@ -958,22 +958,6 @@ async def test_arouter_aretrieve_batch():
         assert mock_aretrieve_batch.call_args.kwargs["api_base"] == "my-custom-base"
 
 
-# ---------------------------------------------------------------------------
-# Batch retrieval has to attribute its tokens to a model group.
-#
-# Batch token usage is accounted on the *retrieve* call, not on create: a
-# provider only reports token counts once the job finishes, so the usage is read
-# off the completed batch's output file during retrieve logging, and that is the
-# spend log row the tokens land on. A batch is retrieved by id, so the request
-# carries no model and the router fans the lookup out across its deployments -
-# the group of the deployment that answered is the only one there is to stamp.
-# Leaving it unset files every batch's tokens under an empty model_group, which
-# is what /global/activity/model groups the spend logs by.
-#
-# The provider is faked at the HTTP boundary, so the retrieve call and the usage
-# accounting that reads the output file both run for real.
-# ---------------------------------------------------------------------------
-
 _BATCH_GROUP = "gemini-batch-group"
 _BATCH_DEPLOYMENT_MODEL = "openai/gpt-4o-mini"
 _BATCH_API_BASE = "http://localhost:4001/v1"